When computers can think, what will they talk about? When they chat with us, and – more intriguingly – with each other?
Until 25 March, Budapest’s C3 gallery and cultural centre hosts a wonderfully atmospheric installation by three Hungarians about just this: robots chatting to each other.
The designers of ‘Smalltalk’ (Zoltan Szegedy-Maszak, Robert Langh, and Marton Fernezelyi) have set up two large screens next to each other, lit from behind by video projectors. As a visitor you enter the darkened room to find both screens facing you, with two green-tinted faces watching you expectantly from the darkness. Both are men in perhaps their late thirties or forties. Both are smoking (after all, this is Eastern Europe) and peering at you from their own darkened rooms. The way the video footage of them has been looped means that – apart from their cigarettes never getting any shorter – they look live – quietly, wearily present.
The effect is pleasingly uncanny. The two chatterbots are watching you and waiting – they want to be set chattering, and are patiently waiting, in their two smoky silences, for you to get them conversing.
From the ceiling on its cord hangs a single thing the size of a pocket calculator, dangling at about chest height.
You, the visitor, tentatively approach this thing and look at it – it is a palm-held computer device with a glowing touch screen. It is encased between two sheets of perspex so that you can only touch the screen and not reach any of its buttons. There is a menu of nine conversation topics and with arrows you can move up and down the menu, or hit a topic.
And so you tap a topic, and the chatterbots start to chatter. The video screens switch to black and white as they start to talk aloud in English, with words appearing in Hungarian on the bottom left and English on the bottom right of each screen. The conversations certainly sound improvised and are different each time you touch on the same topic.
I know this because I tried some topics (such as topic 9: ‘Tell me about the Turing Test’) several times. The two talking heads have a suspicious tendency (at least suspicious to anyone with half an idea of how these programs are put together) to degenerate into quarreling, with the two chatterbots (again and again) accusing each other of being stupid and boring.
But the Turing Test, judging by the delicately spotlit manifesto on one wall, was an important inspiration in creating this exhibit for Szegedy-Maszak, Langh and Fernezelyi. Alan Turing, English mathematician and codebreaker, wrote about this test in 1950 (as the chatterbots never tire of reminding us) as a way of proving computers have become truly intelligent.
Once a true artificial intelligence (AI) had been created, wrote Turing, perhaps by the end of the century (in other words, by about now) we would know because it would be able to pass this simulation test.
The test would involve a computer and a person, each in a separate room, communicating over a text terminal with a human judge. Once the human judge can no longer decide (after some period such as an hour) who is the human and who is the computer – with only the text output to distinguish by – then the computer can be said to have true artificial intelligence.
In other words, computers will be really intelligent when they can bluff.
So when you have two computers – or two computer programs – nattering away with each other, the moot question is: who is fooling whom?
The Hungarian designers, who chose spoken English for the sound because they found the Hungarian word-endings harder to work with, set themselves an extra challenge.
Early chatterbots, such as Weizenbaum’s ELIZA from the 1970s, were simpler to build because they can exploit the imagination of the human interlocutor. The human initiates new lines of discussion, introduces new words – which the chatterbot can then mirror back at the person, embedding them in familiar grammatical structures. In other words, those chatterbots can ask humans intelligent-sounding questions about what the human just said.
ELIZA – a spoof psychoanalyst – would respond to sentences like “I hate my father” with convincing answers like “Why is that? Tell me more about your father.”
Szegedy-Maszak, Langh, and Fernezelyi could not do this – since _both_ sides of the conversation are computer-generated. their technical challenge was trickier. How to keep a real-sounding conversation going when both sides of the chat are randomly assembling grammatical sentences from limited strings of words?
A lovely installation, both soothing and yet thought-provoking, ‘Smalltalk’ makes us ask what we have to hear before we really think there is a mind behind the words? In fact, they undermine the value of Turing’s famous AI test. If a human is part of the conversation, the illusion of intelligence is easier, not harder, to create.
Two chatterbots, as the designers predicted, have a harder time sounding intelligently conversational, than one – the much simpler ELIZA already managed to fool plenty of humans back in the 70s. These two – however witty the scripts they have been given to play with (“I am a program pretending to be a human pretending to be a robot” says the one in glasses at one point) – will fool fewer visitors than simple old ELIZA.
The best part is how the exhibition looks. The green-tinted silent faces, and the black-and-white chatting (rapidly switching back and forward from night to day in the background behind the talking face, depending on when a particular word was taped) make it memorable and mysterious.
If you pass through Budapest go and set them off yourself. 12 noon to 6pm until March 25th, they’re waiting in that room in the Buda Castle District for you, quietly smoking, hoping you’ll pick something for them to talk about.